582
2, MARCH 2018
Bilevel Optimization Framework for Smart
Building-to-Grid Systems
Meysam Razmara, Guna R. Bharati, Mahdi Shahbakhti, Sumit Paudyal, and Rush D. Robinett, III
Abstract—This paper proposes a novel framework suitable
for bilevel optimization in a system of commercial buildings
integrated to smart distribution grid. The proposed optimiza-
tion framework consists of comprehensive mathematical models
of commercial buildings and underlying distribution grid, their
operational constraints, and a bilevel solution approach which
is based on the information exchange between the two levels.
The proposed framework benefits both entities involved in the
building-to-grid (B2G) system, i.e., the operations of the buildings
and the distribution grid. The framework achieves two distinct
objectives: increased load penetration by maximizing the dis-
tribution system load factor and reduced energy cost for the
buildings. This study also proposes a novel B2G index, which is
based on building’s energy cost and nodal load factor, and repre-
sents a metric of combined optimal operations of the commercial
buildings and distribution grid. The usefulness of the proposed
framework is demonstrated in a B2G system that consists of sev-
eral commercial buildings connected to a 33-node distribution test
feeder, where the building parameters are obtained from actual
measurements at an office building at Michigan Technological
University.
Index Terms—Building to grid optimization, smart build-
ing, smart grid, HVAC, model predictive control, hierarchical
optimization, demand response, load shedding.
I. I NTRODUCTION
UILDING sector in the United States accounts for about
B 70% of electricity energy consumption [1], in which
41.4% of energy consumption is directly related to the
space heating, ventilation, and air conditioning (HVAC). Thus,
HVAC systems have a great potential to reduce the energy
usage/cost in buildings. As reported in [2] and [3], smart
control techniques, such as model predictive control (MPC),
can provide significant energy saving. Development of MPC
for optimal control of HVAC systems has been a focus
in [4] and [5], which yields benefits to the operations of
buildings and electric power grid.
Manuscript received June 16, 2015; revised November 21, 2015
and February 21, 2016; accepted April 11, 2016. Date of publication
April 21, 2016; date of current version February 16, 2018. This work was
supported by the Elizabeth and Richard Henes Professorship of Mechanical
Engineering at Michigan Technological University. Paper no. TSG-00687-
2015.
M. Razmara, M. Shahbakhti, and R. D. Robinett, III, are with the
Department of Mechanical Engineering-Engineering Mechanics, Michigan
Technological University, Houghton, MI 49931-1295 USA (e-mail:
mrazmara@mtu.edu; mahdish@mtu.edu; rdrobine@mtu.edu).
G. R. Bharati and S. Paudyal are with the Department of Electrical and
Computer Engineering, Michigan Technological University, Houghton, MI,
USA (e-mail: grbharati@mtu.edu; sumitp@mtu.edu).
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TSG.2016.2557334
1949-3053  c 2016 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/
redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
IEEE TRANSACTIONS ON SMART GRID, VOL. 9, NO.
Smart grid technologies, such as smart appliances and home
area network (HAN), render the electrical loads more man-
ageable and controllable at customers premise, and facilitates
advanced demand side management (DSM) activities [6].
However, existing mathematical models developed for cus-
tomer side DSM mainly consider minimization of electricity
cost [7], [8], which results in increased load at hours with low
electricity price. Increased load at those hours may impact
voltage performance in the distribution grid. In [9], it is
demonstrated that uncontrolled penetration of electric vehicle
loads violates the voltage standards set by the ANSI. Thus, it
is advisable that DSM should be employed considering the
operational requirements of customers as well as the elec-
tric power grid. In DSM activities, distribution utilities are
more concerned about feeder loss minimization, load factor
improvement, reactive power optimization, etc., as part of their
operational objectives [10]–[15]. On the other hand, the objec-
tives of DSM activities at customer level are mainly focused
on comfort maximization and cost minimization. This illus-
trates that the objective set by the grid and customers in DSM
activities could often be conflicting.
In [16] and [17], promising results are obtained from DSM
in reducing customer’s energy price. In [16], pre-cooling
and pre-floating is performed to reduce total electricity cost.
In [17], optimization is performed with a trade-off between
cost and user comfort. In [8], HAN is employed to reshape
the demand profile based on the impact on distribution trans-
former. In DSM models proposed in [8], [16], and [17],
customers obtain benefits by shifting the loads. The mathemat-
ical models used in [8], [16], and [17], consider peak power
information from distribution grid as one of the constraints of
optimization but detailed modeling and operational constraints
of distribution system have not been considered. Thus, for
advanced DSM activities in smart grids, detailed mathematical
modeling considering objectives and operational requirements
of customers and the distribution system with real-time infor-
mation exchange between customer and the grid is crucial;
which is a major focus of the proposed work.
Fig. 1 summarizes past studies related to three research
categories: building, grid, and integrated building-grid
optimization. In the first category, optimal and model
predictive unidirectional building optimization are stud-
ied [2]–[5], [18]–[27]. In this category, building’s HVAC
performance or operational cost is optimized using optimal
or model predictive control technique. The second cate-
gory include studies in which distribution grid operation
is optimized for various operational objectives considering
RAZMARA et al.: BILEVEL OPTIMIZATION FRAMEWORK FOR SMART BUILDING-TO-GRID SYSTEMS
Fig. 1. Summary of past studies on building, grid, and B2G optimization.
aggregated loads [10]–[15]. The third category includes
building-to-grid (B2G) integration in which performance of
buildings including comfort level and energy cost is opti-
mized considering grid side information [7], [16], [28]–[33],
and demonstrates its usefulness for grid services. However,
in the modeling in [7], [16], and [28]–[33] distribution grid
is simplified or completely ignored. Survey results in Fig. 1
reveal that there are extensive works accomplished in the
area of control/optimization of distribution grid and buildings
HVAC systems independently. However, there is little work
done in the area of bi-directional B2G, including mathemat-
ical modeling required for the B2G integration. This work
proposes a bi-directional B2G optimization framework based
on detailed mathematical modeling of a B2G system.
The proposed B2G framework is based on information
exchange between the two levels, i.e., the BEMS and distri-
bution system operation (DSO) control center. The predicted
and optimized load profile of buildings are provided to the
DSO by the BEMSs. Similarly, maximum allowable load pen-
etration information, that ensures feasible grid operations, are
provided to the BEMSs by the DSO.
For the demand response applications discussed in this
work, a bidirectional communication infrastructure is required,
as shown in Fig. 2, between the BEMSs and the DSO. Also,
unidirectional communication links are required between the
BEMSs and the market operator (MO), BEMSs and local
weather station, DSO and control equipment at distribu-
tion level. The required communication infrastructure must
be secure, reliable, and low-cost for autonomous interac-
tions [34], [35]. In this proposed work, the interaction between
the two levels, i.e., the building and grid, are kept at the
minimum; thus, the communication bandwidth usage is spo-
radic. The information that building receives from the grid
control center is the maximum demand limit, which can be
sent once every 15-30 minutes for next couple hours to next
day. Building receives information from utility or MO about
energy price once per day (i.e., the day ahead energy price)
and/or every 5 minutes to one hour (i.e., the real-time energy
price), depending on which energy rate is applied to the build-
ing customers. Building receives weather forecast on daily
583
Fig. 2. Conceptual building to grid communication framework.
basis and/or the hourly weather update. It should be noted that
the information flow depends on the variation of the param-
eters in the building and grid models during the course of
operation, and the type of services the building loads pro-
vide to the grid (such as demand response, regulation). The
communication infrastructure must be flexible to interact with
the existing and future BEMSs and communication protocols
such as Modbus, RS-484, BACnet, etc. [34]. To manage real-
time and bidirectional information exchange cloud computing
can be deployed [35]. VOLTTRON, OpenADR, BEMOSS are
some of the open source platforms that can be utilized for the
implementation of the proposed models at the customer and
grid levels [36]–[38].
This study improves previous single level optimization of
buildings by considering the grid model and builds upon
preliminary results in [39] by the authors. To our knowledge,
it is the first study that proposes a bilevel optimization frame-
work in B2G interaction, which benefits both building and grid
operations. The contribution of this paper is on the develop-
ment of generic hierarchical optimization framework for B2G
system, which is essential for coordinated control of multiple
BEMSs connected to distribution grid for large scale demand
response and other grid level services. More specifically, to
reach this goal this paper:
1) develops a physics-based comprehensive mathematical
model of HVAC system and model predictive controller
(MPC) with the aim of minimizing building’s electricity
costs. The MPC controller works in grid friendly man-
ner, i.e., it communicates with distribution grid control
center and incorporates the constraints set by the grid
controller required for the feasibility of grid operation.
2) develops a detailed and generic mathematical model of
distribution grid that coordinates with several building
controllers to optimize the operation of the power grid.
This paper proposes a novel B2G index that ensures the
benefits at the grid level from the building side optimization,
and develops a coordination algorithm to solve the hierarchical
B2G framework.
584 IEEE TRANSACTIONS ON SMART GRID, VOL. 9, NO. 2, MARCH 2018

The rest of the paper is organized as follows. Section II
presents mathematical models of building and distribution sys-
tem, MPC based optimization model at building level, and
distribution grid optimization model. Section III discusses
methodology for bidirectional optimization and the proposed
B2G index. Section IV describes the building testbed and dis-
area radiated to room i, and Q̇int i denotes the internal heat
generation in room i. Nir is the set of all nodes surrounding
room i.
The disturbance to the model, w(t), is function of neighbor-
ing room temperature, Tkr (t), internal heat generation in rooms
i (t), and radiative heat flux density on walls, Qi (t). The
Q̇int
rad

tribution test feeder. The results of case studies are presented disturbance is given by [40]:
in Section V, and the main conclusions from this work are 
included in Section VI. w(t) = g Tkr (t), Qrad
i (t), Q̇int
i (t) (3)

II. M ATHEMATICAL M ODELING Function g is approximated as a linear function of Tkr (t),

A. Building Components Thermal Modeling
Convection, radiation and conduction are the main heat
transfer ways in buildings which are time varying. We use
a common building modeling approach known as nodal
approach to model building’s construction and materials with
electrical components such as resistors, capacitors, and current
sources [2], [40]. For this purpose, buildings are considered
as graphs containing walls and rooms as nodes. p represents
number of nodes, q out of p nodes are rooms, and the remain-
ing p − q nodes are walls. i = 1, . . . , q numbers are assigned
to each room. The following equation determines the heat
transfer of the walls [40]:
⎛ ⎞
w
dTi,j 1 ⎜  T r − Tw
k i,j rad ⎟
= w⎝ wi,j + ri,j γi,j Aw
i,j Qi,j ⎠ (1)
dt Ci,j R mass flow rate and its temperature for each zone is shown with
k∈N w k
i,j
where Ti,j w is the temperature of the wall between room i
and room j, Ci,j w is the heat capacity of the wall between
r
room i and j, Tk is the temperature of adjacent rooms to wall
wi,j . Thermal resistance between the center-line of wall (node
w sured using a vent mass flow meter. Density, heat capacity,
wi,j ) and the neighboring node k is denoted with Rk i,j . γi,j and
Aw
i,j are radiation heat absorption coefficient and area of wall
between room i and j, respectively. Wall identifier is shown by
ri,j which is equal to 0 for internal walls, and equal to 1 for
peripheral walls. Qrad
i,j represents the radiative heat flux density
on wall (i, j) while Ni,j w is the set of all neighboring nodes to
node wi,j .
The following equation governs the temperature of the ith
room [40]:
⎛ varying known input to the model. Given that the air mass
i (t) and Q̇i (t). Thus, the disturbance is w = aTk (t) +
Qrad int r
bQi (t) + cQ̇i (t).
rad int
Heat transfer of each wall and room equations and distur-
bance form the system dynamics of building are represented
in state-space form by:
ẋ(t) = f (x(t), u(t), w(t), t)
y(t) = Cx(t) (4)
where x(t) ∈ Rn is the state vector. State vector includes
nodes’ temperature in the nodal network. y(t) ∈ Rm is the
output vector of the system presenting thermal zones’ temper-
ature. Thermal zone is a section of a building that is required
to meet certain comfort level. In this paper, each room is con-
sidered as a thermal zone. The input vector which is the air
u(t) ∈ Rl×m . l is the number of inputs to each thermal zone
(i.e., air mass flow and supply air temperature). C is the out-
put matrix with proper dimension to return output, y(t), from
states.
In our testbed, HVAC air mass flow rate is directly mea-
trasnsmissivity and etc. are determined using standard engi-
neering thermodynamics tables. There are unknown (e.g., wall
heat capacitance and outside air convection coefficient) that
are determined for the testbed using unscented Kalman fil-
ter (UKF) techniques presented in [2].
Equation (4) describes the nonlinear time evolution of the
system. The supply air temperature (i.e., input to the system)
is multiplied by the air mass flow, which is the other time-

dTir 1 ⎝  Tk − Tir flow rate is constant during day, we can consider the linearized
= r + πi,k τkw Awin Q
i,k i
rad
form of the system. We use Euler’s discretization method to
dt Ci Ri,k
k∈Nir discretize the state update equation (4) for controller design.
⎞
 
+ ṁri cpavg Tis − Tir + Q̇inti
⎠ (2)
where Cir and ṁri denote the heat capacity and air mass flow
into or out of the room i, respectively. Tk is the temperature of
surrounding node k to room i. cpavg denotes the average specific
heat capacity of air and Tis is the temperature of the supply air
to room i. πi,j is window identifier which is equal to 0 if there
is no wall between room i and j, otherwise equal to 1. Awin i,j
is the total area of window between room i and surrounding
room j, τi,j
w is the transmissivity of glass of window between
room i and j, Qradi is the radiative heat flux density per unit
The state-space model is given by:
xk+1 = Axk + Buk + Ewk (5)
where wk is the disturbance vector at instant k and E is the
disturbance coefficient matrix with appropriate dimension.
Fig. 3 represents the experimental validation of the building
thermal model for few days in winter. Fig. 3(a) compares the
simulated and measured room temperature, and Fig. 3(b) illus-
trates the heat-pump power based on the difference between
the room temperature and the measured supply air temper-
ature. Details of HVAC heat-pump system are discussed in
Section IV.
RAZMARA et al.: BILEVEL OPTIMIZATION FRAMEWORK FOR SMART BUILDING-TO-GRID SYSTEMS 585

subject to: xt+k+1|t = Axt+k|t + But+k|t + Edt+k|t (8b)

yt+k|t = Cxt+k|t (8c)
U t+k|t ≤ ut+k|t ≤ U (8d)
δU ≤ ut+k+1|t − ut+k|t ≤ δU (8e)
T t+k|t − εt+k|t ≤ yt+k|t ≤ T t+k|t + ε t+k|t (8f)
εt+k|t , ε t+k|t ≥ 0 (8g)

where (8b) and (8c) form building’s state model, (8d) and (8e)
are constraints on input i.e., supply air temperature, (8f) is
output constraint on temperature of room and (8g) represents
the constraint on slack variables. Constraints (8b) and (8d)
should hold for all k = 0, 1, . . . , N − 1, and N is the predic-
Fig. 3. Experimental validation of the building thermal model for a sample
room/zone. tion horizon. In Section V-F, effect of N on simulation results
will be discussed. (8c), (8e), (8f), and (8g) should hold for
B. Building Optimization Model-I all k = 1, 2, . . . , N. Ut = [ut|t , ut+1|t , . . . , ut+N−1|t ] represents
The main goal of optimization is to minimize energy cost control inputs vector and ut+1|t is the estimated value of ut+1
by the building’s HVAC system. For this purpose, both HVAC at time t,  t = [εt+1|t , . . . , ε t+N|t ] and  t = [εt+1|t , . . . , ε t+N|t ]
energy usage and time varying electricity price are taken into are the slack variables. Slack variables are added to ensure
account. Energy usage is characterized by the energy index Ie feasibility of optimal control problem. In equation (8a) there
which is defined as [2]: is penalty (ρ) for slack variables. Therefore, by choosing large
values for ρ, the optimizer enforces the slack variables to take

24
small values and does not violate the temperature constraints.
Ite = PBt t (6a)
ρ characterizes the flexibility of a building to provide ser-
t=1
f vices at the grid level.  is not a pre-determined value in
PBt = Pht + Pct + Pt + Pot (6b) the optimization; however, by choosing large enough values
where PBt is the active power consumption of building. Ph is for ρ, deviation from comfort bound (i.e., ) is guaranteed to
heating power, Pc , and P f are cooling power and fan power, be minimum. In equation (8a), lower values for ρ leads to
respectively. We denote all buildings’ other loads with Po higher flexibility for grid level services; however, this results
which includes lighting and appliances loads. Heating, cooling in occupants’ discomfort [2]. Equation (8f) defines the rela-
and fan power consumption are determined by [2]: tionship between the room temperature and the slack variable
  value.  is dynamic pricing of electricity which is consid-
Pht = ṁri (t)cp,air Tth − Ti,t
r
(7a) ered to be independent of power consumption by the building
 r  loads. yt+k|t is the vector of thermal zone temperature, dt+k|t is
Pt = ṁi (t)cp,air Ti,t − Tt
c r c
(7b)
 r 3 the disturbance load, and T t+k|t and T t+k|t for k = 1, · · · , N
f
Pt = λ ṁi (7c) are the lower and upper bounds on the zone comfort level,
respectively. U t+k|t and U are the lower and upper limits on
where T h and T c are the supply air temperatures in the heating
the supply air temperature delivered by the heat-pump of the
mode and the cooling mode, respectively and these variables
HVAC system, respectively. Operational limit on maximum
are the system control inputs. Equations (7a) and (7b) rep-
supply air temperature is not time varying, therefore time
resent the relationship between the HVAC heat-pump power
invariant constraint U is used. δU and δU are limitation on rate
consumption and the temperature difference. λ [W.s3 .kg−3 ] is
of change of supply air temperature according to dynamics of
the coefficient of fan which defines the cubic relation between
heat-pump condenser.
power required and mass flow rate. Since air mass flow rate is
not a control input in the state model, and it is constant during
the daytime, fan power, Pf , does not affect optimization result. C. Building Optimization Model-II
The temperature difference between supply air and the room
air is proportional to electrical energy consumption. Hence, Objective of this optimization model is the same as (8).
the building controller keeps the room temperature within the However, an additional constraint is used as a feedback from
comfort levels such that energy cost is minimized. We use distribution grid, which is given as:
the proposed objective function in equation (8a) to minimize
energy cost. In addition, soft constraints (i.e., ) are imple- PBt ≤ Pvar
t (9)
mented to guarantee feasibility of optimal solution at all times.
The following optimization problem is being solved at each where Pvar
t is the maximum allowable building active power
time step t, and cumulative cost is calculated from t to tmax : penetration in the distribution grid, which will be discussed
 t   in Section II-D. Equations (6), (7a)-(7c), (8b)-(8g), and (9)
max
 
min It  + ρ | t |1 + | t |1
e T
(8a) define equality and inequality constraints of the optimization
Ut ,¯ , model.
t
586 IEEE TRANSACTIONS ON SMART GRID, VOL. 9, NO. 2, MARCH 2018

D. Distribution Grid Optimization Model-I Inequality constraints of the distribution grid optimiza-
The objective of this optimization model is to determine tion model include limits of voltage as per ANSI standards,
maximum additional building loads that can be connected limits on capacitor and LTC positions, limits on ItF , and
at different nodes of the distribution system. This infor- other limits such as feeder capacity, transformer capacity, etc.
mation is used as a feedback to the building optimization Mathematically, the inequality constraints can be written as,
model-II described in Section II-C, which represents infor- V ≤ Vtn ≤ V (14a)
mation exchange between the two levels of the bilevel opti- n
mization framework. Mathematically, objective function can 0≤ Ctapnt ≤ Ctap , ∀ Ctapnt ∈ I (14b)
be written as: Ttap ≤ Ttapmt ≤ Ttap, ∀ Ttapt ∈ I
m
(14c)
t 
max 
var,nb 0≤ ItF ≤ I
F
(14d)
max Pt (10)
t=1 nb∈n In equations (14a, 14b, 14c, & 14d) overbar and underbar
where n represents nodes and nb represents nodes where denote the maximum and minimum value for variable, respec-
buildings are connected. tively. Equation (14a) ensures that the voltage limits are within
Equality constraints of the grid optimization model include the standard set by ANSI. Equations (14b) and (14c) are the
component modeling and voltage/current balance equations. limits on tap positions in transformer and capacitor banks,
Distribution feeders, transformers, and voltage regulators with which take integer numbers only.
load tap changers (LTCs) are modeled using ABCD param-
eters as described in [41]. Constant power, constant cur- E. Distribution Grid Optimization Model-II
rent, and constant impedance loads are considered at each Objective in this optimization model is to maximize the
node. Mathematically, these equality constraints can be rep- system load factor. Mathematically, this objective function is
resented as: written as [42] and [43]:
n
m

Vt a bm Vtn+1  t 
= m (11a) max
Itsen,m c dm Itres,m t=1 Pt
max (15)
tmax |Pt |∞
Itres,m−1 = Itsen,m + ItZ,n + ItI,n
+ ItP,n + ItC,n + Itvar,n (11b) where, Pt total active power load connected to the distribution
grid at hour t is calculated as:
Z n ItC,n = Vtn Ctapnt (11c)
  Z,n ∗  
where m represents feeder branches, Itres,m−1
and are Itsen,m Pt = Vtn It + ItI,n + ItP,n + Pvar,nb
t (16)
receiving and sending end currents on branch m − 1 and n nb
m, respectively. For branches with LTC, am = 1+s·Ttap
1
m, Equations (11), (12), and (14) represent the equality and
t
bj = cj = 0 and dj = a−1 j where s represents voltage inequality constraints. Load factor in (15) is improved by con-
change (p.u.) with one step change of transformer tap position trolling LTC and switched capacitor banks. For grid power
and Ttapm n
t represents LTC tap positions, Z is the impedance
flow analysis and distribution grid optimization model-II,
of single capacitor in a bank at nominal power and voltage. Ptvar,nb = PB,nb
t , which is obtained from (6).
ItP,n , ItI,n , ItZ,n , and ItC,n are the currents from the constant
power, constant current, constant impedance loads, and capac- F. B2G Index
itor banks connected at node n, respectively. Vtn represents In the hierarchical approach to solve the proposed optimiza-
nodal voltage and Ctapnt represents number of capacitor banks tion models, the lower level is concerned about the electricity
switched on. costs in buildings, while the upper level is concerned in maxi-
Base loads in the distribution feeder are modeled using sum mizing the system load factor. A worthwhile B2G optimization
of constant power, constant current, and constant impedance should maintain an appropriate compromise between low elec-
loads. Building loads are additional power to the grid and mod- tricity cost in buildings and high load factor in the distribution
eled as constant power load. Building load current in terms system. To account for this, we define a new B2G index (IB2G )
of building power can be calculated using: to assess the performance of the proposed B2G optimization
∗ Pvar,nb framework, which allows reduction in the electricity costs and
Vtnb Itvar,nb = t ∠φ (12) improvement of system load factor starting at the building level
cosφ
with the use of nodal load factor. The B2G index is defined as:
where φ is the power factor angle of the building load. An
additional equation is needed to ensure that building load pene- Lf−1
    
tration is allowed fairly in the distribution grid. Thus, a fairness  
index (ItF ) is defined to ensure fair distribution of building  e T tmax PtB,nb 
∞
loads as: IB2G = α It  + β t (17)
max B,nb
P
t=1 t
Ptvar,nb
ItF = (13) where α and β are the weight factors for energy price (Ie T )
PtE,nb
and nodal load factor (Lf ), respectively. The smaller the IB2G
where, PtE,nb is the base load in nodes nb. index, the better the performance for the B2G optimization.
RAZMARA et al.: BILEVEL OPTIMIZATION FRAMEWORK FOR SMART BUILDING-TO-GRID SYSTEMS 587

The ratio of weight factors, α to β, determines the importance

of energy cost compared to the system load factor. α and β are
the design parameters in the B2G optimization problem and
can be adjusted by the building and grid operators, depending
on whether the benefits from the building side or the ben-
efits from the grid side are more desirable. Thus, selection
of proper values of α and β for optimal operation system
requires sensitivity analysis of the building-grid system under
consideration.
In the case studies, IB2G is used as an alternative objec-
tive function to (8a) in building optimization models. Note
that (8a) does not account for grid’s objective; while the pro-
posed B2G index considers the grid’s objective in the building
optimization model with consideration of the nodal load factor.

III. S OLUTION M ETHOD

A solution method, shown in Fig. 4, is proposed for the
bilevel B2G optimization framework. The proposed solution
method consists of information exchange between the two lev-
els, i.e., the commercial buildings and the grid control center;
thus, it is bidirectional in nature. The B2G optimization frame-
work consists of two optimization models for the demand
side (i.e., building systems) and two optimization models for
the supply side (i.e., distribution grid). These optimization
procedures include:
• Building Optimization Model-I: Building energy cost
minimization;
• Distribution Grid Optimization Model-I: Maximizing
penetration of building loads in the distribution grid;
• Building Optimization Model-II: Minimizing building
electricity cost considering constraints from grid side (9);
• Distribution Grid Optimization Model-II: Maximizing
load factor of distribution grid.
The solution procedure begins with the demand side opti-
mization with an objective to reduce electricity cost for the
next day (label A in Fig. 4). At this stage, the optimized load
profiles of each of the commercial buildings are made avail-
able to the distribution grid control center. A distribution grid
power flow analysis is carried out to confirm the feasibility of
building load profiles. The feasibility test checks various oper-
ational constraints of the distribution grid including voltage
limits as per ANSI standards defined by (14a)-(14c). Violation
of operational constraints makes the building loads in-feasible
for grid operation. In case the building loads are feasible for
the grid (label B in Fig. 4), tap positions of capacitor banks and
transformers are calculated using distribution grid optimization
model-II, which maximizes the system load factor.
In case of in-feasibility of the building loads from grid side,
B,ib
the maximum allowable load, Pt , is obtained from the distri-
B,ib
bution grid optimization model-I. Pt is used as a feedback to
the BMS system (label D in Fig. 4) to solve the building opti-
mization model-II. If the updated optimization is feasible, then
the optimized load profile is sent back to the distribution opti-
mization model-II to maximize load factor (label F in Fig. 4).
Otherwise, other options like changing building’s tempera-
ture requirements (label E in Fig. 4) becomes necessary.
Next, building optimization model-II is solved by considering
Fig. 4. Proposed bidirectional B2G optimization flowchart.
the feedback from the grid as additional constraint. Then, the
resulting building load profile information is sent to the distri-
bution grid control center, which is used to solve distribution
optimization model-II.
Change of the control variables, i.e., LTC and capacitor
banks may result in different node voltages from the two distri-
bution grid optimization models, which in turn can impact the
voltage dependent loads and make the grid optimization prob-
lems in-feasible. Thus, depending on the types of load used
in the distribution grid optimization model, the whole solution
process may need to reiterate. An alternate approach to han-
dle this issue is to consider a small unused margin of power
in the maximum allowable loads in (10) that can accommo-
date change in power consumption of loads due to the voltage
change. However, this change will not affect the feasibility of
HVAC system since they can be considered as constant power
load [44].
IV. T EST S YSTEMS
A. Building Testbed
A commercial three story Lakeshore Center building
with an area of 61, 500 f t2 at Michigan Technological
University (MTU), Houghton, Michigan, is considered as the
testbed. The testbed is equipped with ground-source heat-
pumps (GSHP) to provide required energy for heating and
cooling. GSHPs transfer geothermal energy from the ground
to rooms. GSHPs are HVAC energy efficient technologies with
high coefficient of performance (COP). The GSHPs in this
study have a nominal COP of 3.2.
588
Fig. 5. Schematic of the testbed with a ground-source heat-pump. Two
temperature sensors are used to measure the average room air temperature.
Each room is equipped with a GSHP to maintain com-
fort of occupants. Fig. 5 shows the schematic of the building
with GSHPs connected to the distribution grid at node #10.
Temperature of each zone in Lakeshore Center building is
measured using a temperature sensor from the BMS with
accuracy of ±0.2◦ C. HVAC system in the testbed consumes
up to 44% of electricity in winter, while the rest of elec-
tricity consumption is attributed to lighting, computers, and
office appliances. Historical data of energy consumption of an
office building at MTU is used to compute the share of energy
consumption of major loads in the testbed.
B. Distribution Test Feeder
A standard single phase 12.66 kV, 33-node distribution
feeder [11] is considered for the case studies. The standard test
system is modified by adding a regulator and capacitor banks
to account for the control equipment available in distribution
grids.
As shown in Fig. 5, a transformer is equipped with voltage
regulator (connected between 10th and 11th nodes in the test
systems), which maintains the voltage at node 11 within ±10%
of the nominal value. A 32-step regulator is considered, and a
single tap change represents voltage of 0.00625 p.u. Capacitor
banks are connected at 8th and 16th nodes of the feeder and
are modeled as constant impedance load capable of delivering
reactive power to the grid. Identical five units of capacitors
with 10 kVAr each are considered, which are represented as
switched capacitors with tap positions from 0 to 5.
It is assumed that smart buildings are connected at four
arbitrary nodes, i.e., nodes 2, 10, 18, and 31. Each building
is considered to have twenty identical zones with similar load
profiles and with the same temperature requirements.
V. R ESULTS
The building model is implemented in MATLAB and
YALMIP [45] toolbox is used to provide a symbolic syntax
to formulate the problem and interfaces with the optimization
IEEE TRANSACTIONS ON SMART GRID, VOL. 9, NO. 2, MARCH 2018
solver. Distribution system optimization model is developed in
GAMS [46]. The accuracy of the building model is validated
using actual measurements collected from an office building
at Michigan Tech [2]. The distribution grid model is validated
using the results in [11].
This work considers the weather forecast and energy
price as the inputs required for the building and grid opti-
mization models. Thus, electricity dynamic pricing available
from Midcontinent Independent System Operator (MISO) is
used [47]. Weather forecast information available from a local
weather station is used. The constraints on the room tem-
perature are set based on ASHRAE standards [48]. Other
building and grid parameters used in the simulation can be
found in [20] and [11], respectively.
The optimization results are compared with the non-
optimized rule-based controller (RBC). We use same t = 1
hour time step to have a fair comparison among B2G optimiz-
ers in this study. Given the slow thermal behavior of the testbed
building and the type of demand response applications, one
hour time step is sufficient to capture the dynamics. Depending
on building type, a shorter time step may be required.
For the case study, five buildings each having twenty zones
are connected at node 2 of the distribution grid, similar three
buildings are connected at node 10, four buildings are con-
nected at node 18, and seven buildings are connected at
node 31. In this paper, the optimization results are only pre-
sented for the buildings at node #10 (see Fig. 5). The nodal
load factor is presented for the node #10 as a node repre-
senting B2G system. In addition, system level load factors are
also determined for the distribution grid system to compare
the operation of different B2G optimizers.
As mentioned in the Section I, HVAC systems account for a
large amount of load in buildings. Therefore, in this study, we
focus on controlling the HVAC load of the buildings which is
considered to be dispatchable. Other loads in buildings includ-
ing lighting, appliances, and base loads in the distribution grid
are considered to be non-dispatchable. For MPC simulation, a
prediction horizon of N = 24 is used with time step of 1 hour.
Here, we introduce three different optimization methods and
compare them with the RBC.
A. Unoptimized Rule Based Control
In this approach, when temperature goes below the lower-
bound temperature, the HVAC RBC keeps the heat-pump
compressor on for the duration of t. In the next time
step, the RBC checks the room temperature again and deter-
mines whether the room temperature is within the comfort
bounds. The monthly electricity cost of the building and
energy consumption along with other B2G metrics are listed
in TABLE I.
B. Building/Grid Side Optimizations
The results of uncoordinated building side and grid side
optimization are presented here. Fig. 6 shows the results of a
building MPC controller using building optimization model-I.
The main objective in the building-side optimization is to
minimize building electricity costs. Fig. 6(a) demonstrates the
RAZMARA et al.: BILEVEL OPTIMIZATION FRAMEWORK FOR SMART BUILDING-TO-GRID SYSTEMS
TABLE I
P ERFORMANCE OF F OUR D IFFERENT B2G O PTIMIZATION T ECHNIQUES , C OMPARED TO THE U NOPTIMIZED (RBC) C ASE S TUDY
Fig. 6. Building-side optimization: (a) Control input and resulting temper-
ature profile for the building MPC controller, (b) Buildings’ load profile
including HVAC load, lighting load and appliances load, (c) Buildings’
total load at node #10 and base load, (d) Buildings’ electricity load versus
maximum feasible load for node #10.
temperature profile of the room, and Fig. 6 (b) shows load dis-
tribution of buildings connected to node #10 including lighting
load, appliances (computers, elevators, refrigerators, etc.) and
buildings’ HVAC load. Building load distribution is based on
simulation results carried out for Lakeshore Center building at
Michigan Tech using Carrier HAP software. Fig. 6(c) shows
load distribution on node #10. Fig. 6(d) illustrates the building
electricity load and the maximum feasible load for node #10
which is obtained from distribution optimization model-I block
shown in Fig. 4. In Fig. 6(d), the importance of grid-wise
optimization for buildings is emphasized since the optimized
building loads violate the maximum allowable loads by the
grid. Note that in Fig. 6, the power and supply temperature
peaks at unusual time is due to the pre-heating of the HVAC
system when the electricity price is cheaper. The dynamic pric-
ing and result HVAC cost are shown in Fig. 7. The monthly
building electricity cost is listed in Table I. Compared to the
RBC, building-side optimization results in 26% cost saving
and 16% energy saving.
589
Fig. 7. Building-side Optimization: dynamic pricing versus hourly cost for
the building.
If the objective is to maximize load factor in the distribution
grid (i.e., the grid-side optimization), the nodal load factor is
significantly increased to 0.84, while the building’s energy cost
increases.
C. Bidirectional Optimization
The proposed B2G methodology in Section III is applied
for the case study similar to that in Section V-B. Building
optimization results are illustrated in Fig. 8 and summarized in
Table I. The results show the B2G optimizer satisfies the build-
ing comfort levels, while keeping the building load under the
maximum load allowed by the distribution grid. In addition,
compared to the unidirectional (demand side) optimization, the
system level load factor is increased from 0.81 to 0.84. IB2G
index also depicts a significant improvement as seen in Table I.
Bidirectional optimizer offers 25% cost saving compared to the
unoptimized case while energy cost saving is only dropped
by 1% compared to unidirectional (building side). Thus, this
case study demonstrates that with the grid constraints, the cost
of operation of buildings increases, but it ensures a feasible
operation of the grid.
D. Bidirectional Optimization Using IB2G Index
IB2G index provides a way to control/formulate the impor-
tance of building benefits versus grid benefits. The results of
bidirectional optimization using IB2G index are shown in Fig. 9
and summarized in the Table I. As expected, the new opti-
mization approach leads to the best result in terms of IB2G .
The smaller value for IB2G , the better performance for the
B2G system. Using IB2G in the building objective function
helps to improve building/nodal load factor and consequently
590 IEEE TRANSACTIONS ON SMART GRID, VOL. 9, NO. 2, MARCH 2018

TABLE II
P ERFORMANCE OF T WO D IFFERENT B2G R EAL -T IME O PTIMIZATION T ECHNIQUES , C OMPARED TO THE U NOPTIMIZED (RBC) C ASE S TUDY

Fig. 8. Bidirectional optimization: (a) Control input and room temperature Fig. 9. Bidirectional optimization based on IB2G index: (a) Control input
profile for the MPC controller, (b) Building load vs maximum feasible load. and resulting temperature profile, (b) Building load versus maximum feasible
load.
better system level load factor, while minimizing the building
energy costs. As a result, IB2G is better in the new bidirectional
optimization, compared to the bidirectional optimization with
objective 8a. 0.5 and 0.54 are chosen for α and β, respectively,
to provide comparable weight to building and grid. Selection
of α and β can be another optimization problem which is
outside the scope of this work. In this case study, the cost
of electricity is decreased by 24% with respect to the base
case and both nodal and system load factors are significantly
improved compared to the base case.
If α in IB2G in the bidirectional optimization is chosen to
be zero, the bidirectional optimization becomes unidirectional
which mainly satisfies the grid benefits (Grid unidirectional
in Table I). As expected, this leads to the best load factor
compared to other four cases in Table I. But, the buildings’
electricity cost is increased by 21% compared to the bidi-
rectional IB2G case. Overall, bidirectional optimization using Fig. 10. Voltage at selective nodes. Fig. (a) shows the voltage profile
IB2G provides an optimization platform in which an operator of unidirectional optimization and Fig. (b) illustrates the voltage profile of
bidirectional optimization (using IB2G index).
can easily decide the desirable compromise for the benefits of
customers and the grid. By using the proposed IB2G index as
the optimization cost function, the temperature boundaries are
met based on the ASHRAE standards. F. Hour-Ahead vs Day-Ahead Optimization
In the case studies in Table I, the prediction horizon N is
E. Impact on Voltage Performance considered to be 24. It means that the optimization algorithm
Permissible voltage range, as specified in ANSI C84.1, is has prediction of weather forecast and dynamic pricing for the
one of the standards to maintain power quality in distribution next 24 hours. If the weather forecast or prediction of electric-
feeders. Fig. 10(a) shows the voltage at few selected nodes ity cost (dynamic pricing) is not available or not accurate for
for unidirectional optimization (building-side) which are vio- that period, the MPC algorithm cannot find the optimal solu-
lated in certain hours, and (b) illustrates that the optimization tion for the next 24 hours. Here, we show two examples that
model with IB2G index where voltages are maintained within the dynamic pricing and the weather forecast are available
the limits. for the next hour in Table II. Simulations for two proposed
RAZMARA et al.: BILEVEL OPTIMIZATION FRAMEWORK FOR SMART BUILDING-TO-GRID SYSTEMS
Fig. 11. (a) weather forecast of three sample days in Fall, Winter, and Spring
using measured data at Michigan Tech testbed and (b) Dynamic pricing for
Michigan, Illinois, and Minnesota obtained from [47].
methods, bidirectional using (8a) and bidirectional using IB2G ,
are carried out to demonstrate the effect of prediction hori-
zon on the cost saving and IB2G index. Results show that
B2G metrics deteriorate by reducing the prediction horizon
but the proposed algorithm still offers benefits for both grid
and buildings. Using the real-time B2G optimization, build-
ing electricity cost drops up to 20% and building load factor
increases over 0.1 compared to the baseline unoptimized case.
G. Impact of Dynamic Pricing and Seasonal Weather
The savings on electricity price and improvement on sys-
tem load factor greatly depend on various factors including
dynamic pricing, weather, flexibility of commercial build-
ing’s loads, building’s temperature set-points, available control
equipment in the building and distribution grid, and accu-
racy/availability of the forecasts. The impact of dynamic
pricing and weather variations on the customer’s cost savings
and system load factor are also studied.
Fig. 11(a) shows the weather condition for three sample
days in Fall, Winter, and Spring in our testbed. Fig. 11 (b)
illustrates three different dynamic energy prices at nodes in
Michigan, Illinois and Minnesota. Based on the weather and
energy price data, two sets of simulations are carried out.
Fig. 12(a) shows the effect of weather on the building’s elec-
tricity cost and system load factor for the three seasons by
keeping the dynamic energy price same (Michigan node).
Fig. 12(b) demonstrates the effect of dynamic pricing on the
electricity cost saving and the system load factor improve-
ment using same weather forecast (Fall). During the Winter,
33% saving in electricity cost and 8% improvement on sys-
tem load factor. However, as weather gets warmer, the cost
saving decreases. For instance, in Spring, only 6% decrease
in electricity bill compared to RBC is observed. The reason
for less saving in Spring compared to Winter is that the cold
outside weather in Winter makes HVAC running more often
which provides more opportunities to save energy costs by
shifting the load. Less constraints (i.e., heating requirement)
591
Fig. 12. (a) shows the weather effect on the building’s electricity bill and
system load factor, and (b) illustrates the effect of dynamic pricing on the
electricity bill and system load factor for the conditions shown in Fig. 11.
An RBC structure similar to that in Table I is used as a baseline to calculate
saving percentages.
on the HVAC controller in Spring/Fall compared to Winter
provides more flexibility for load factor optimization. Thus
the percentage of increase in the B2G system’s LF is more in
Fall/Spring compared to that in the Winter. Finally, the results
in Fig. 12 (b) show that the proposed B2G optimization pro-
vides cost saving and system LF increase for all the three
dynamic pricing profiles studied.
H. Computational Cost
The problem formulation is hierarchical in nature, which
allows each BEMS to solve its own building optimiza-
tion model. The developed building optimization model in
Section II is linear programming in nature, thus the com-
putational tasks at the building level are not challenging for
the demand response applications discussed in this work. At
the grid level, the nature of the problem is non-linear and
LTCs and cap banks add integer variables in the model. This
poses inherent computational challenges associated with the
large scale non-convex optimization of distribution grid [49].
However, for the size of the system considered in this paper,
computational complexity was not an issue. Simulation time
for bidirectional optimization discussed in this paper on an
INTEL Core i5, 3.2 GHz CPU desktop computer is less
than one minute for building’s unidirectional optimization and
around 5 minutes for optimization of both building and grid
models. The entire hierarchical optimization problem takes
maximum of two iterations to solve. First, optimal load pro-
files at each interval (i.e., each hour in this work) from the
buildings are sent to grid control center to determine feasi-
bility of the grid operation. In case of infeasibility, maximum
feasible load profiles are sent to each building controllers.
Computational challenges associated with large practi-
cal size grid can be reduced by using methods of con-
vexification [50], distributed approaches [14], and heuristic
approaches [10]. The solution time desired for the proposed
models depend on the B2G applications sought. For example,
592
voltage regulation needs to be tackle in a few minutes, and
load shifting in the order of hours [51]. For the type of B2G
applications proposed in this work, solution time of 5 minutes
and a coarse time interval of one hour suffice.
I. Benefits, Challenges, and Limitations
This work provides a generic mathematical framework to
optimally coordinate building loads and grid level assets,
which is useful for near-term energy usage planning and/or
near-real time dispatch of building loads. This opens up oppor-
tunities to deploy multiple distributed building loads for grid
level applications such as demand response, load follow-
ing, and regulation services. The framework ensures that, in
the demand dispatch process, the operational constraints and
interests of the grid level and customer level energy manage-
ment activities are honored; thus providing benefits to both the
entities involved.
The case studies clearly demonstrated the benefits of the
proposed framework to the building and grid operations.
However, the major challenges in large scale deployment of
B2G integration are: (1) infrastructural challenges including
interoperability of tools at building and customer levels, band-
width limitation, compatibility of system to handle different
data with different resolution and with different communi-
cation standards, and (2) mechanism barriers including lack
of financial models for costumer motivation, accurate pred-
ication of weather and energy price, computational issues
for grid optimization for real-time applications in practical-
sized systems, and scale-ability of the control and optimization
algorithms [51].
Use of dynamic pricing (day ahead or real-time) at
small residential and commercial buildings has already
begun [52], [53]. With the widespread implementation of
dynamic pricing for small customers, and with energy
management systems at customer’s premise, the power peak
in distribution systems will shift to the hours with least
expensive energy prices. Thus, the utility needs to devise
demand limits for the customers for the feasible grid opera-
tion. However, with the demand limits, energy savings of the
customer will reduce. An appropriate incentive/compensation
mechanism need to take place [54], where utilities may
compensate the customers for the reduction in the savings.
However, the compensation may not be necessary for all
cases, since the case studies demonstrated that the optimized
energy costs of the buildings are reduced compared to the
RBC even with the grid constraints. Utilities need to devise
subscription plans where the customers agree that the BEMSs
honor the demand limits sent by the DSO and penalties for
noncompliance. The another approach to solve this issue is
the use of distribution location marginal price (DLMP) [55].
If appropriate method for setting DLMP can be designed,
which considers the customer’s expected response for the
low energy prices, the impact of low energy price on peak
demand can be eliminated to a certain extent.
VI. C ONCLUSION
This study develops a bidirectional B2G optimization frame-
work using an experimentally validated building thermal
IEEE TRANSACTIONS ON SMART GRID, VOL. 9, NO. 2, MARCH 2018
model and a mathematical model for the distribution grid. The
building optimization model is based on disturbance and heat
transfer phenomena, and its model parameters are obtained
from actual measurements collected from an office building
at Michigan Technological University. The distribution grid
model is comprised of modeling of typical distribution sys-
tem components including feeders, transformers, and control
equipment such as capacitor banks and transformer load tap
changers. In the proposed bidirectional optimization model,
the objective is to minimize energy cost for the demand side
and to maximize load factor for the grid. To account for con-
flicting interests of the BEMS and distribution operators in
the bidirectional optimization, a novel B2G index is devel-
oped based on building energy cost and nodal load factor.
Based on the provided case studies, it can be concluded that
the developed bidirectional optimization framework can reduce
commercial buildings’ monthly electricity costs by 25% in
Winter, compared to the unoptimized rule-based control of
the building loads, while improving the system load factor.
However, the savings obtained in energy price and improve-
ment on system load factor greatly depend on various factors
including energy price, flexibility of commercial building
loads, customers’ preferences, available control equipment in a
distribution grid, accuracy/availability of weather forecast and
dynamic pricing prediction, etc. Nevertheless, the developed
bidirectional optimization framework certainly offers benefits
to the customers and the utilities in B2G integration.
ACKNOWLEDGMENT
Special thanks to G. Kaurala from MTU Energy
Management Office for help in collecting the building exper-
imental data.
R EFERENCES
[1] (Dec. 10, 2014). Building Energy Data Book. [Online]. Available:
http://buildingsdatabook.eren.doe.gov
[2] M. Maasoumy, M. Razmara, M. Shahbakhti, and A. S. Vincentelli,
“Handling model uncertainty in model predictive control for energy
efficient buildings,” J. Energy Build., vol. 77, pp. 377–392, Jul. 2014.
[3] S. Privara et al., “Building modeling as a crucial part for building
predictive control,” J. Energy Build., vol. 56, pp. 8–22, Jan. 2013.
[4] J. Ma, S. J. Qin, B. Li, and T. Salsbury, “Economic model predictive
control for building energy systems,” in Proc. IEEE PES Innov. Smart
Grid Technol., 2011, pp. 1–6.
[5] M. Pčolka, E. Žáčeková, R. Robinett, S. Čelikovský, and M. Šebek,
“Economical nonlinear model predictive control for building climate
control,” in Proc. Amer. Control Conf. (ACC), Portland, OR, USA, 2014,
pp. 418–423.
[6] A. Brooks, E. Lu, D. Reicher, C. Spirakis, and B. Weihl, “Demand
dispatch,” IEEE Power Energy Mag., vol. 8, no. 3, pp. 20–29,
May/Jun. 2010.
[7] M. C. Bozchalui, S. A. Hashmi, H. Hassen, C. A. Cañizares, and
K. Bhattacharya, “Optimal operation of residential energy hubs in smart
grids,” IEEE Trans. Smart Grid, vol. 3, no. 4, pp. 1755–1766, Dec. 2012.
[8] S. Shao, M. Pipattanasomporn, and S. Rahman, “Demand response as
a load shaping tool in an intelligent grid with electric vehicles,” IEEE
Trans. Smart Grid, vol. 2, no. 4, pp. 624–631, Dec. 2011.
[9] J. Taylor, A. Maitra, M. Alexander, D. Brooks, and M. Duvall,
“Evaluations of plug-in electric vehicle distribution system impacts,” in
Proc. IEEE Power Energy Soc. Gen. Meeting, Minneapolis, MN, USA,
2010, pp. 1–6.
[10] S. Paudyal, C. A. Cañizares, and K. Bhattacharya, “Optimal operation of
distribution feeders in smart grids,” IEEE Trans. Ind. Electron., vol. 58,
no. 10, pp. 4495–4503, Oct. 2011.
[11] S. K. Goswami and S. K. Basu, “A new algorithm for the reconfiguration
of distribution feeders for loss minimization,” IEEE Trans. Power Del.,
vol. 7, no. 3, pp. 1484–1491, Jul. 1992.
RAZMARA et al.: BILEVEL OPTIMIZATION FRAMEWORK FOR SMART BUILDING-TO-GRID SYSTEMS
[12] Y. Deng, X. Ren, C. Zhao, and D. Zhao, “A heuristic and algorithmic
combined approach for reactive power optimization with time-varying
load demand in distribution systems,” IEEE Trans. Power Syst., vol. 17,
no. 4, pp. 1068–1072, Nov. 2002.
[13] I. Roytelman, B. K. Wee, and R. L. Lugtu, “Volt/var control algorithm
for modern distribution management system,” IEEE Trans. Power Syst.,
vol. 10, no. 3, pp. 1454–1460, Aug. 1995.
[14] E. Dall’Anese, H. Zhu, and G. B. Giannakis, “Distributed optimal power
flow for smart microgrids,” IEEE Trans. Smart Grid, vol. 4, no. 3,
pp. 1464–1475, Sep. 2013.
[15] M. B. Liu, C. A. Cañizares, and W. Huang, “Reactive power and voltage
control in distribution systems with limited switching operations,” IEEE
Trans. Power Syst., vol. 24, no. 2, pp. 889–899, May 2009.
[16] M. Avci, M. Erkoc, and S. S. Asfour, “Residential HVAC load con-
trol strategy in real-time electricity pricing environment,” in Proc.
Energytech, Cleveland, OH, USA, May 2012, pp. 1–6.
[17] D. T. Nguyen and L. B. Le, “Joint optimization of electric vehicle and
home energy scheduling considering user comfort preference,” IEEE
Trans. Smart Grid, vol. 5, no. 1, pp. 188–199, Jan. 2014.
[18] Y. Yan et al., “Adaptive optimal control model for building cooling and
heating sources,” J. Energy Build., vol. 40, no. 8, pp. 1394–1401, 2008.
[19] M. Maasoumy, A. Pinto, and A. Sangiovanni-Vincentelli, “Model-
based hierarchical optimal control design for HVAC systems,” in Proc.
ASME Dyn. Syst. Control Conf. (DSCC), Arlington, TX, USA, 2011,
pp. 271–278.
[20] M. Razmara, M. Maasoumy, M. Shahbakhti, and R. D. Robinett, III,
“Optimal exergy control of building HVAC system,” J. Appl. Energy,
vol. 156, pp. 555–565, Oct. 2015.
[21] X. Xu, S. Wang, and G. Huang, “Robust MPC for temperature control of
air-conditioning systems concerning on constraints and multitype uncer-
tainties,” J. Build. Services Eng. Res. Technol., vol. 31, no. 1, pp. 39–55,
2010.
[22] M. Razmara, M. Maasoumy, M. Shahbakhti, and R. D. Robinett, III,
“Exergy-based model predictive control for building HVAC systems,”
in Proc. Amer. Control Conf. (ACC), Chicago, IL, USA, 2015,
pp. 1677–1682.
[23] G. Huang, S. Wang, and X. Xu, “A robust model predictive control strat-
egy for improving the control performance of air-conditioning systems,”
J. Energy Convers. Manag., vol. 50, no. 10, pp. 2650–2658, 2009.
[24] F. Oldewurtel et al., “Energy efficient building climate control using
stochastic model predictive control and weather predictions,” in
Proc. Amer. Control Conf. (ACC), Baltimore, MD, USA, Jun. 2010,
pp. 5100–5105.
[25] Y. Ma, S. Vichik, and F. Borrelli, “Fast stochastic MPC with optimal
risk allocation applied to building control systems,” in Proc. IEEE 51st
Annu. Conf. Decis. Control, Dec. 2012, pp. 7559–7564.
[26] P.-D. Moroşan, R. Bourdais, D. Dumur, and J. Buisson, “Building
temperature regulation using a distributed model predictive control,”
J. Energy Build., vol. 42, no. 9, pp. 1445–1452, 2010.
[27] Y. Ma, G. Anderson, and F. Borrelli, “A distributed predictive control
approach to building temperature regulation,” in Proc. Amer. Control
Conf. (ACC), San Francisco, CA, USA, 2011, pp. 2089–2094.
[28] M. Maasoumy and A. Sangiovanni-Vincentelli, “Buildings to grid inte-
gration: A dynamic contract approach,” in Proc. IEEE/ACM Int. Conf.
Comput.-Aided Design, Austin, TX, USA, 2015, pp. 473–478.
[29] T. Gamauf, T. Leber, K. Pollhammer, and F. Kupzog, “A generalized
load management gateway coupling smart buildings to the grid,” in Proc.
AFRICON Conf., Livingstone, Zambia, 2011, pp. 1–5.
[30] P. Palensky and D. Dietrich, “Demand side management: Demand
response, intelligent energy systems, and smart loads,” IEEE Trans. Ind.
Informat., vol. 7, no. 3, pp. 381–388, Aug. 2011.
[31] X. Xue and W. Shengwei, “Interactive building load management for
smart grid,” in Proc. IEEE Power Eng. Autom. Conf., Wuhan, China,
2012, pp. 1–5.
[32] A. Saha et al., “A home energy management algorithm in a smart house
integrated with renewable energy,” in Proc. IEEE Innov. Smart Grid
Technol. Conf. Europe, Istanbul, Turkey, 2014, pp. 1–6.
[33] B. Jin, P. Nuzzo, M. Maasoumy, Y. Zhou, and
A. Sangiovanni-Vincentelli, “A contract-based framework for integrated
demand response management in smart grids,” in Proc. 2nd ACM Int.
Conf. Embedded Syst. Energy Efficient Built Environ., Seoul, South
Korea, 2015, pp. 167–176.
[34] J. Hagerman, G. Hernandez, A. Nicholls, and N. Foster, “Buildings-to-
grid technical opportunities: Introduction and vision,” U.S. Dept. Energy,
Energy Efficient and Renewable Energy (EERE), Washington, DC, USA,
Tech. Rep., 2014.
[35] S. Bera, S. Misra, and J. P. C. Rodrigues, “Cloud computing applications
for smart grid: A survey,” IEEE Trans. Parallel Distrib. Syst., vol. 26,
no. 5, pp. 1477–1494, May 2015.
[36] J. Haack, B. Akyol, B. Carpenter, C. Tews, and L. Foglesong, “Volttron:
An agent platform for the smart grid,” in Proc. Int. Conf. Auton. Agents
Multi-Agent Syst., St. Paul, MN, USA, May 2013, pp. 1367–1368.
593
[37] W. Khamphanchai et al., “Conceptual architecture of building energy
management open source software (BEMOSS),” in Proc. Innov. Smart
Grid Technol. Conf. Europe, Istanbul, Turkey, 2014, pp. 1–6.
[38] G. Ghatikar and R. Bienert, “Smart grid standards and systems interop-
erability: A precedent with OpenADR,” in Proc. Grid Interop Forum,
Phoenix, AZ, USA, 2011, pp. 1–9.
[39] M. Razmara, G. R. Bharati, M. Shahbakhti, S. Paudyal, and
R. D. Robinett, III, “Bidirectional optimal operation of smart building-
to-grid systems,” in Proc. Amer. Control Conf. (ACC), Chicago, IL, USA,
2015, pp. 288–293.
[40] M. Maasoumy, M. Razmara, M. Shahbakhti, and A. S. Vincentelli,
“Selecting building predictive control based on model uncertainty,”
in Proc. Amer. Control Conf. (ACC), Portland, OR, USA, 2014,
pp. 404–411.
[41] W. H. Kersting, Distribution System Modeling and Analysis. Boca Raton,
FL, USA: CRC Press, 2012, Chs. 3–6.
[42] E. Sortomme, M. M. Hindi, S. D. J. MacPherson, and S. S. Venkata,
“Coordinated charging of plug-in hybrid electric vehicles to minimize
distribution system losses,” IEEE Trans. Smart Grid, vol. 2, no. 1,
pp. 198–205, Mar. 2011.
[43] R. F. Chang and C. N. Lu, “Feeder reconfiguration for load fac-
tor improvement,” in Proc. Power Eng. Soc. Winter Meeting, vol. 2.
New York, NY, USA, 2002, pp. 980–984.
[44] A. Bokhari et al., “Experimental determination of the zip coefficients
for modern residential, commercial, and industrial loads,” IEEE Trans.
Power Del., vol. 29, no. 3, pp. 1372–1381, Jun. 2014.
[45] J. Löfberg, “YALMIP: A toolbox for modeling and optimization in
MATLAB,” in Proc. IEEE Int. Symp. Comput. Aided Control Syst.
Design, Taipei, Taiwan, 2004, pp. 284–289.
[46] GAMS Development Corporation. (Dec. 10, 2014). General
Algebric Modeling System, Release 24.2.1. [Online]. Available:
http://www.gams.com
[47] MISO. (Sep. 3, 2014). Midcontinent Independent System Operator.
[Online]. Available: https://www.misoenergy.org/MarketsOperations/
Prices/Pages/Prices.aspx
[48] ASHRAE, “Ventilation for acceptable indoor air quality,” Amer. Soc.
Heating, Refrigerating Air-Conditioning Eng. Inc, Atlanta, GA, USA,
Standard 62.1-2004, 2004.
[49] S. Paudyal, C. A. Cañizares, and K. Bhattacharya, “Three-phase distri-
bution OPF in smart grids: Optimality versus computational burden,” in
Proc. Innov. Smart Grid Technol. Europe, Manchester, U.K., Dec. 2011,
pp. 1–7.
[50] S. H. Low, “Convex relaxation of optimal power flow—Part I:
Formulations and equivalence,” IEEE Trans. Control Netw. Syst., vol. 1,
no. 1, pp. 15–27, Mar. 2014.
[51] T. Kuruganti and M. Brambley, “Building-to-grid technical opportu-
nity: From the information and communications technology perspective,”
DOE Energy Efficiency & Renew. Energy, Washington, DC, USA, Tech.
Rep. DOE/EE-1054, Feb. 2016.
[52] Ameren. Power Smart Pricing; A Smart Electricity Rate From
Ameren Illinois. Accessed on Jun. 3, 2015. [Online]. Available:
https://www.powersmartpricing.org/
[53] The COMED Residential Real-Time Pricing Program
Guide, Commonwealth Edison (ComEd), Chicago, IL,
USA. Accessed on Jun. 3, 2015. [Online]. Available:
https://www.comed.com/documents/customer-service/
rates-pricing/real-time-pricing/rrtpprogramguide.pdf
[54] Z. Wang, “Compensation methods for demand response,” M.S. thesis,
Dept. Elect. Comput. Eng., Univ. Kentucky, Lexington, KY, USA, 2015.
[55] R. Li, Q. Wu, and S. S. Oren, “Distribution locational marginal pricing
for optimal electric vehicle charging management,” IEEE Trans. Power
Syst., vol. 29, no. 1, pp. 203–211, Jan. 2014.
Meysam Razmara is currently pursuing the Ph.D. degree in mechanical
engineering from Michigan Technological University, Houghton, MI, USA.
Guna R. Bharati is currently pursuing the Ph.D. degree in electrical
engineering from Michigan Technological University, Houghton, MI, USA.
Mahdi Shahbakhti is an Assistant Professor of Mechanical Engineering at
Michigan Technological University, Houghton, MI, USA.
Sumit Paudyal is an Assistant Professor of Electrical Engineering at
Michigan Technological University, Houghton, MI, USA.
Rush D. Robinett, III, is the Richard and Elizabeth Henes Chair Professor of
Mechanical Engineering and the Co-Director of the Agile and Interconnected
Microgrids Center, Michigan Technological University, Houghton, MI, USA.